Principal AI Systems Architect

About the role
This is the person who builds that architecture. As Principal AI Systems Architect you own Mimoid's long-term technical architecture and drive it from today's shipping product to the full platform vision. You will design and stand up the shipbuilding-native stack the company compounds on, and you will be hands-on — containerizing services, deploying into the cloud, and prototyping the hard parts, not just drawing diagrams. You work directly with the CEO and CTO and set the standards the rest of engineering builds against.

Mimoid's platform is organized around four capabilities you will architect for the future of the domain. Each is described below alongside the core skills it demands:

Multi-modal information extraction — design systems that ingest heterogeneous engineering inputs — documents, technical drawings, specifications, CAD exports, spreadsheets, and field photography — and transform them into a single canonical bill of materials and schedule, with source, confidence, and full provenance attached to every value.

Skills: document AI and computer vision, ML-driven data extraction pipelines, schema and ontology design, and data-quality/confidence modeling.

Domain-specific agent orchestration — architect a roster of specialized AI agents (e.g., bill-of-materials transformation, work packaging, scheduling, change-impact, workflow advisory, quality, and vision-based progress tracking), each running a context model write-back loop behind human-in-the-loop review gates.

Skills: multi-agent system design, agent runtimes and tool-calling, workflow orchestration, and human-in-the-loop control patterns.

Private, bounded model training — build focused domain agents that can run on local or isolated infrastructure and improve over time per site, without depending on external services.

Skills: model fine-tuning (e.g., LoRA), private and edge inference, rules-engine design, and continuous model improvement.

Knowledge-learning platform — design a bi-temporal manufacturing knowledge graph and a domain ontology that compound build over build, connected by a privacy-preserving, consent-gated network that lets learning transfer across sites.

Skills: knowledge-graph and ontology engineering, bi-temporal data modeling, and consent- and privacy-aware data sharing.

What you'll do

• Own the end-to-end target architecture and the phased roadmap to reach it — evolving the Phase-1 modular monolith (FastAPI API + Celery worker, managed data) into service-ready boundaries that extract into microservices and a service mesh only when a named trigger forces it.
• Architect and operate the cloud + deployment foundation: containerize services, run them in the cloud, and design per-project isolation as a hard gate. Own CI/CD, secrets/encryption, observability, and cost-per-build instrumentation.
• Design the knowledge-graph core: a property-graph database (e.g., Neo4j) paired with a relational + vector store (PostgreSQL/pgvector) for hybrid retrieval, bi-temporal and change-impact edges, provenance on every value, and the domain ontology — plus the context service that grounds and token-budgets every agent call.
• Architect the agent runtime — orchestration, tool calling, human-in-the-loop gates, and trace/replay/cost telemetry — and a versioned, eval-gated prompt registry.
• Own the model-serving abstraction: route every call by classification and complexity across a tiered set of managed models (small medium large), CPU-servable local embeddings, a deterministic mock provider for CI, and a local open-weight + LoRA fallback.
• Design the continuous-learning plane: capture evaluate (eval harness, calibration, trust tiers) curate adapt (LoRA / rule update) promote (shadow canary production with auto-rollback), plus an anonymized, consent-gated, privacy-preserving cross-site learning network.
• Keep the numeric decision path deterministic and auditable — scheduling (critical-path methods) and flow optimization (constraint solvers such as OR-Tools CP-SAT) run with no LLM in the loop.
• Write the ADRs and reference architecture; make build-vs-buy, open-weight-vs-proprietary, and coupling decisions; own FinOps for LLM and compute spend.

Required qualifications

• 10+ years in software/systems engineering, including 4+ years owning production system architecture end-to-end.
• Cloud & deployment (hands-on): deep experience on a major cloud platform — containerization (Docker), container runtime/orchestration, infrastructure-as-code, CI/CD, and running managed data services (PostgreSQL, object storage, Redis) in production. You can stand the system up yourself, not just specify it.
• Knowledge graph: hands-on graph experience — Neo4j (or a comparable property-graph engine), property-graph/ontology modeling, and graph-backed retrieval; comfort pairing a graph with a vector store (pgvector or equivalent) for hybrid retrieval.
• AI / agents: architecting LLM and multi-agent systems — RAG, embeddings/reranking, agent runtimes and tool-calling, prompt/version registries, evals, and LLMOps/observability.
• Model serving: experience routing across models/providers and reasoning about cost/latency/quality tradeoffs.
• Data & correctness: relational + vector data modeling, provenance/lineage, and keeping deterministic compute separate from probabilistic components.
• Fluency architecting polyglot systems (Python/FastAPI + TypeScript/Node) and writing clear ADRs and reference architectures.

Preferred qualifications

• Experience taking an early-stage platform to a compliant, multi-tenant production posture (e.g., SOC 2, ISO 27001).
• Hands-on with private, on-premise, or air-gapped inference environments.
• Local/open-weight model fine-tuning (LoRA) and evaluation-gated model promotion.
• Operations-research / optimization exposure (CP-SAT, critical-path scheduling).
• Domain background in PLM/PDM/BOM systems or complex high-mix, low-volume manufacturing